A Fault Diagnosis Method Based on Adaptive Noise Reduction Convolutional Neural Network

Abstract

The present invention relates to a fault diagnosis method based on an adaptive noise reduction convolutional neural network. The method includes the following steps: detecting the vibration signals of mechanical equipment, forming a one-dimensional vibration data set, performing grayscale mapping on the vibration data set to obtain a grayscale image database; inputting the grayscale image database into the adaptive noise reduction convolutional neural network for training to obtain its adaptive noise reduction neural network model; and using the trained adaptive noise reduction neural network model for fault diagnosis. Compared with the prior art, the present invention constructs an adaptive noise reduction convolutional neural network, which contains an adaptive filter, enabling the convolutional neural network to adaptively perform fault diagnosis under high noise, and having the advantages of strong anti-noise ability and high accuracy.

Description

Technical Field

[0001] The present invention relates to the field of mechanical equipment fault diagnosis, and particularly to a fault diagnosis method based on an adaptive noise reduction convolutional neural network. Background Art

[0002] With the improvement of the intelligence level of mechanical equipment, the fault diagnosis method based on the traditional model, also known as the model-driven fault diagnosis method, has gradually been replaced by the data-driven fault diagnosis method and has been widely applied to a large number of mechanical equipment, such as bearings, gearboxes, transmission shafts, etc.

[0003] The convolutional neural network is a commonly used deep learning method in the field of data-driven fault diagnosis, and a large number of research results have been obtained in the fault diagnosis of various mechanical equipment. The data-driven fault diagnosis method based on the convolutional neural network mainly uses the equipment signal data set to train the convolutional neural network to obtain a neural network model, and uses the trained convolutional neural network model to diagnose the faults of the equipment. However, the fault diagnosis algorithms represented by the convolutional neural network often ignore the actual working scenario, that is, the operation of the machine often generates many additional noise interferences, such as the resonance between mechanical structures. Although a large number of auxiliary devices and higher-precision sensors are used to filter the signals, the interference of noise in the data still exists.

[0004] However, the performance of its convolutional neural network has a strong relationship with the noise level of the data. On the one hand, when the noise contained in the data becomes larger, the traditional convolutional neural network cannot effectively filter the noise signal, and the accuracy of the convolutional neural network model for fault diagnosis often drops significantly. On the other hand, the traditional fault diagnosis method based on the convolutional neural network cannot adapt to the fault diagnosis in different operating environments. And if the noise reduction is simply carried out by humans or auxiliary devices, it often requires a lot of manual effort and has uncertainty. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a fault diagnosis method based on an adaptive noise reduction convolutional neural network.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A fault diagnosis method based on an adaptive noise reduction convolutional neural network includes the following steps:

[0008] Step S1: Detect the vibration signal of the mechanical equipment and form a one-dimensional vibration data set;

[0009] Step S2: Perform grayscale processing on the vibration data set to obtain a grayscale image database;

[0010] Step S3: Input the grayscale image database into the constructed adaptive noise reduction convolutional neural network for training to obtain its adaptive noise reduction neural network model. The adaptive noise reduction neural network model includes an adaptive filtering convolutional layer, a global average pooling layer, and a fully connected layer. The adaptive filtering convolutional layer consists of two CNN layers and an adaptive filter, and the threshold of the adaptive filter is adjusted by an automatic threshold setting module.

[0011] Step S4: Perform fault diagnosis through the trained noise reduction neural network model.

[0012] Further, step S1 specifically includes:

[0013] S11: Collect the vibration signals of the mechanical equipment to form a one-dimensional labeled vibration dataset L(i), where i represents the i-th data point in the one-dimensional vibration data.

[0014] S12: Select M*M one-dimensional vibration data from the vibration dataset, and substitute each data point in M*M into the grayscale calculation formula to obtain P(j,k), where P(j,k) represents the size of the pixel at the j-th row and k-th column in the grayscale image.

[0015] S13: Use P(j,k) to form an M*M grayscale image.

[0016] Further, the expression of the grayscale calculation formula is:

[0017]

[0018] Among them, the function round represents the rounding function to ensure that P(j,k) is an integer between 0 and 255.

[0019] Further, the adaptive noise reduction neural network model includes 4 adaptive filtering convolutional layers, 1 global average pooling layer, and 1 fully connected layer.

[0020] Further, the automatic threshold setting module includes 1 absolute value layer, 1 global pooling layer, and 2 one-dimensional CNN layers.

[0021] Further, the automatic threshold setting module outputs an adaptive threshold to the adaptive filter to achieve automatic threshold setting. The data filtered by the adaptive filter is added to the input image data, and the result is then input into the next adaptive filtering convolutional layer.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] An adaptive filter is added to the convolutional neural network in the present invention to achieve filtering of noise signals. At the same time, the present invention uses two one-dimensional convolutional layers to achieve adaptive setting of the threshold of the adaptive filter, eliminating the interference caused by manual threshold setting, thereby achieving fault diagnosis with strong anti-noise ability and high accuracy. Description of the Drawings

[0024] Figure 1 It is a schematic flow chart of the present invention.

[0025] Figure 2 It is a schematic structural diagram of the adaptive noise reduction convolutional neural network.

[0026] Figure 3 It is a schematic structural diagram of the automatic threshold setting module. Detailed Implementation Manner

[0027] The present invention will be described in detail below with reference to the drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives the detailed implementation manner and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0028] As Figure 1 shown, this embodiment provides a fault diagnosis method based on an adaptive noise reduction convolutional neural network, which specifically includes the following steps:

[0029] Step S1: Detect the vibration signal of the mechanical equipment and form a one-dimensional vibration data set.

[0030] Step S2: Perform grayscale processing on the vibration data set to obtain a grayscale image database.

[0031] Step S3: Construct an adaptive noise reduction convolutional neural network including 4 adaptive filter convolutional layers, 1 global average pooling layer, and 1 fully connected layer. The adaptive filter convolutional layer is composed of two CNN layers and an adaptive filter, and the threshold of the adaptive filter is adjusted by the automatic threshold setting module. Input the grayscale image database into the constructed adaptive noise reduction convolutional neural network for training to obtain its adaptive noise reduction neural network model.

[0032] Step S4: Use the trained noise reduction neural network model for fault diagnosis.

[0033] I. Conversion of One-Dimensional Vibration Data Set into Grayscale Image

[0034] Detect the vibration signal of the mechanical equipment and form a one-dimensional vibration data set. Perform grayscale processing on the vibration data set to obtain a grayscale image database. The processing flow is as follows.

[0035] (1) Collect the vibration signals of mechanical equipment to form a one-dimensional vibration dataset with labels. L(i), where i ∈ [0, n], represents the i-th data point in the one-dimensional vibration data.

[0036] (2) Select M * M one-dimensional vibration data from the vibration dataset. For each data point (L(i), where i ∈ [0, M 2 ) in M * M, substitute it into formula (1) to obtain P(j, k).

[0037]

[0038] Among them, the function round represents the rounding function to ensure that P(j, k) is an integer from 0 to 255.

[0039] (3) Use P(j, k) to form a grayscale image of M * M. P(j, k) represents the size of the pixel at the j-th row and k-th column in the grayscale image;

[0040] Adopt the above method to form multiple grayscale images and then form a grayscale image database for subsequent model training.

[0041] II. Train an adaptive noise reduction convolutional neural network using the grayscale image database

[0042] (1) Construct an adaptive noise reduction convolutional neural network;

[0043] The adaptive noise reduction convolutional neural network includes 4 adaptive filtering convolutional layers, 1 global average pooling layer, and 1 fully connected layer. Among them, the adaptive filtering convolutional layer consists of two CNN layers and an adaptive filter. The threshold of the adaptive filter is adjusted by an automatic threshold setting module, and the automatic threshold setting module includes 1 absolute value layer, 1 global pooling layer, and 2 one-dimensional CNN layers. Its network structure is as Figure 2 shown.

[0044] (2) Input the grayscale image into the constructed adaptive noise reduction convolutional neural network to train the adaptive noise reduction neural network model. The specific process is as follows:

[0045] An image with a size of M * M enters the first adaptive filtering convolutional layer. First, the feature data x is extracted through two layers of CNN layers jk (the size of the pixel at the j-th row and k-th column in the grayscale image). After extraction, the feature data x jk is then filtered using the adaptive filter. The adaptive threshold of the adaptive filter is generated by the automatic threshold setting module, as Figure 3 shown. In the automatic threshold setting module, the feature data x jk first passes through the absolute value layer to take the absolute value, and the absolute value data passes through the global average pooling layer. The calculation of the global average pooling is shown in formula (2),

[0046]

[0047] Where W and H represent the width and height of the entire image respectively, and g(x) represents the output of the global average pooling layer.

[0048] The global average pooling layer compresses the grayscale image into a single number. This one-dimensional number serves as the input to the one-dimensional CNN of layer 2. The first layer of the one-dimensional CNN is activated by the ReLu function, and the second layer of the one-dimensional CNN is activated by the Sigmoid function. After being processed by the two layers of one-dimensional CNN, the adaptive threshold of the adaptive filter is finally obtained, and the threshold calculation is as shown in formula (3).

[0049]

[0050] Where σ(·) represents the Sigmoid function, as shown in formula (4). The Sigmoid function can ensure that the output value is within the range of 0 to 1.

[0051]

[0052] Represents two layers of one-dimensional convolution operations, where K1 and K2 represent the convolution kernels of each layer. Represents the convolution operation, as shown in formula (5).

[0053]

[0054] Where ReLu is a commonly used activation function, as shown in formula (6). In formula (6): Represents the output value after the convolution operation; Is The value after passing through the Relu function.

[0055]

[0056] The automatic threshold setting module outputs the adaptive threshold to the adaptive filter to achieve automatic threshold setting. The data filtered by the adaptive filter is then added to the input image data, and the resulting data is input into the next adaptive filtering convolution layer. In this way, 4 adaptive filtering convolution layers are used for processing in sequence. The processed data is classified through the global average pooling layer and the fully connected layer, and the classification result is finally output. The above entire network structure constitutes an adaptive noise reduction neural network model.

[0057] The adaptive noise reduction neural network model is trained with the grayscale image database, and the adaptive noise reduction neural network model is continuously updated until the trained adaptive noise reduction convolutional neural network model is finally obtained.

[0058] III. Fault Diagnosis Using an Adaptive Denoising Convolutional Neural Network

[0059] Input the fault data into the adaptive denoising convolutional neural network model to achieve fault diagnosis.

[0060] IV. Numerical Example Analysis

[0061] For better explanation, the proposed method is implemented and verified using the dataset of a wind turbine gearbox as follows:

[0062] Collect the vibration data of the wind turbine gearbox to form a one-dimensional vibration dataset, and perform grayscale mapping on the vibration dataset to obtain a grayscale image database. The processing flow is as follows.

[0063] (1) Collect the vibration signals of the mechanical equipment to form a one-dimensional vibration dataset with labels. L(i), i ∈ [0, n] represents the i-th data point in the one-dimensional vibration data.

[0064] (2) Select M * M one-dimensional vibration data from the vibration dataset, and substitute each data point (L(i), i ∈ [0, M 2 ) into formula (1) to obtain P(j, k).

[0065] Among them, the function round represents the rounding function to ensure that P(j, k) is an integer from 0 to 255.

[0066] (3) Use P(j, k) to form an M * M grayscale image, where P(j, k) represents the size of the pixel at the j-th row and k-th column in the grayscale image; use this method to form multiple grayscale images and then form a grayscale image database for subsequent model training.

[0067] (4) The specific process of the adaptive denoising convolutional neural network is as follows:

[0068] An image with a size of M * M enters the first adaptive filtering convolutional layer, and first passes through two CNN layers to extract the feature data x jk (the size of the pixel at the j-th row and k-th column in the grayscale image). After extraction, the feature data x jk is then filtered using an adaptive filter, where the adaptive threshold of the adaptive filter is generated by the automatic threshold setting module. In the automatic threshold setting module, the feature data x jk first passes through an absolute value layer to take the absolute value, and the absolute value data passes through a global average pooling layer. The calculation of the global average pooling is shown in formula (2), where W and H respectively represent the width and height of the entire image, and g(x) represents the output of the global average pooling layer.

[0069] The global average pooling layer compresses the grayscale image into a single number. This one-dimensional number serves as the input to the one-dimensional CNN in layer 2. The first layer of the one-dimensional CNN is activated by the ReLu function, and the second layer of the one-dimensional CNN is activated by the Sigmoid function. After being processed by the two layers of the one-dimensional CNN, the adaptive threshold of the adaptive filter is finally obtained, and the threshold calculation is as shown in formula (3).

[0070] Where σ(·) represents the Sigmoid function, as shown in formula (4). The Sigmoid function can ensure that the output value is within the range of 0 to 1.

[0071] represents two layers of one-dimensional convolution operations, where K1 and K2 represent the convolution kernels of each layer, represents the convolution operation, as shown in formula (5).

[0072] Where ReLu is a commonly used activation function, as shown in formula (6). In formula (6): represents the output value after the convolution operation; is the value after passing through the Relu function.

[0073] The automatic threshold setting module outputs the adaptive threshold to the adaptive filter to achieve the automatic setting of the threshold. The data filtered by the adaptive filter is then added to the input image data, and the resulting data is input into the next convolutional layer with an adaptive filter. In this way, 4 adaptive filter convolutional layers are used for processing in sequence. The processed data is classified through the global average pooling layer and the fully connected layer, and finally the classification result is output.

[0074] The performance of the adaptive noise reduction neural network is compared with 4 other convolutional neural networks, namely AlexNet, LeNet, ResNet, and DRSN-CW. The results are shown in Table 1.

[0075] Table 1 Performance of each method under different noises

[0076]

[0077] According to the test results, the fault diagnosis method conceived in the present invention has the following advantages. (1) An adaptive filter is added to the convolutional neural network in this method to achieve the filtering of noise signals. (2) By using two layers of one-dimensional convolutional layers, the adaptive setting of the threshold of the adaptive filter is realized, eliminating the interference caused by manual threshold setting.

[0078] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field according to the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A fault diagnosis method based on an adaptive noise reduction convolutional neural network, characterized in that, Including the following steps: Step S1: Detect the vibration signals of the mechanical equipment and form a one-dimensional vibration data set; Step S2: Perform grayscale mapping on the vibration data set to obtain a grayscale image database; Step S3: Input the grayscale image database into the constructed adaptive noise reduction convolutional neural network for training to obtain its adaptive noise reduction neural network model; the adaptive noise reduction neural network model includes an adaptive filtering convolutional layer, a global average pooling layer, and a fully connected layer, where the adaptive filtering convolutional layer consists of two CNN layers and an adaptive filter, and the threshold of the adaptive filter is adjusted by an automatic threshold setting module; Step S4: Perform fault diagnosis through the trained noise reduction neural network model; The adaptive noise reduction neural network model includes 4 adaptive filtering convolutional layers, 1 global average pooling layer, and 1 fully connected layer; The automatic threshold setting module includes 1 absolute value layer, 1 global pooling layer, and 2 one-dimensional CNN layers; The automatic threshold setting module outputs an adaptive threshold to the adaptive filter to achieve automatic threshold setting; the data filtered by the adaptive filter is added to the input image data, and the result is input to the next adaptive filtering convolutional layer.

2. The fault diagnosis method based on an adaptive noise reduction convolutional neural network according to claim 1, wherein, Step S1 specifically includes: S11: Collect the vibration signals of the mechanical equipment and form a one-dimensional labeled vibration data set L(i), where i represents the i-th data point in the one-dimensional vibration data; S12: Select M*M one-dimensional vibration data from the vibration data set, and substitute each data point in M*M into the grayscale calculation formula to obtain P(j,k), where P(j,k) represents the size of the pixel at the j-th row and k-th column in the grayscale image; S13: Use P(j,k) to form an M*M grayscale image.

3. The fault diagnosis method based on an adaptive noise reduction convolutional neural network according to claim 2, wherein The expression of the grayscale calculation formula is: where the function round represents the rounding function to ensure that P(j,k) is an integer between 0 and 255.

Images